Design And Implementation Of An Anomaly Detection Model For Multivariate Time Series Data

Multivariate time series anomaly detection is a data analysis technique used to identify abnormal values or mutation events in multivariate time series data.In today’s digital age,with the continuous development of industrial systems and data acquisition technologies,applying multivariate time series anomaly detection techniques can timely detect potential problems and avoid unnecessary economic losses.In the traditional manufacturing field,statistical methods are usually used to model each variable in multivariate time series data separately.However,this method is too dependent on the knowledge and experience of engineers,and cannot fully utilize the spatial relationships between multiple variables.Therefore,this paper deeply studies multivariate time series data,and designs a robust multivariate time series anomaly detection model that can fully explore the spatio-temporal relationships of data,which is applied to the semiconductor wafer manufacturing field.This paper first proposes a spatial feature extractor based on graph attention networks to capture the spatial relationships between data variables.The spatial feature extractor encodes each variable into a feature vector and uses these embedding representations to construct a variable relationship graph,and then extracts spatial features through the graph.Then,the paper designs a time feature extractor based on self-attention mechanism to capture the temporal relationships between data.To obtain a representation of the overall spatio-temporal relationship,the paper designs a spatio-temporal feature fusion network based on position encoding and Transformer to combine the obtained spatial and temporal relationships.In addition,the paper adopts a joint training strategy,which inputs the obtained overall spatio-temporal relationship representations into a reconstruction model based on LSTM and VAE and a prediction model based on Bi GRU,respectively.Joint training can simultaneously utilize the ability of the reconstruction model to capture the data distribution of the entire time window and the modeling ability of the prediction model for individual time point data.To address the balance between tasks in joint training,the paper designs a dynamic loss function.Finally,the paper designs an anomaly scoring mechanism that integrates the results of the reconstruction model and the prediction model to obtain more comprehensive and reliable inferences.The model designed in this paper is fully experimentally designed and verified on two public industrial datasets and one real semiconductor wafer production equipment dataset.The experimental results show that the designed model can effectively capture the spatio-temporal relationships between multivariate data.This paper also proves the effectiveness and necessity of each module in this model through multiple detailed ablation experiments from multiple perspectives.Currently,the multivariate time series data anomaly detection model designed in this paper has been applied to the semiconductor wafer manufacturing field.